This paper studies bitcoin address usage, which is assumed to be hidden via address pseudonyms. Transaction anonymity is ensured by means of bitcoin addresses, leading to abuse for illegitimate purposes, e.g., payments of illegal drugs, ransom, fraud, and money laundering. Although all the transactions are available in the bitcoin system, it is not trivial to determine the usage of addresses. This work aims to estimate typical usages of bitcoin transactions based on transaction features. With the decision tree learning algorithm, the proposed algorithm classifies a set of unknown addresses into seven classes; provider addresses of three services for mining pool, Bitcoin ATM, and dark websites; and user addresses of four services for mining Bitcoin ATM, dark websites, exchange, and a bulletin board system. The experimental results reveal some useful characteristics of bitcoin traffic, including statistics of frequency, amount of value, and significant transaction features.